Group Activity Recognition in Visual Data Using Deep Learning Framework

Group activity recognition, a subset of humanactivity recognition, has garnered significant research focus due to its crucial role in comprehending videos. Particularly, group activity recognition (GAR) involves identifying activities performed by a collective of individuals. The primary hurdle in this domain lies in grasping the interplay amongst a group within a scene and how it evolves over time. GAR aims to identify the overall activity of a group, holding substantial potential across applications like sports/surveillance video analysis, social activity comprehension, and video search and retrieval. Effectively comprehending scenes with multiple individuals demands a model that not only characterizes each actor's individual actions in context but also deduces their collective activities. While tackling a video clip, this challenge entails not only recognizing individual actions but also uncovering scene information and the collaborative dynamics among people. Recent advancements have witnessed the emergence of deep network-based techniques and graphical models for addressing group activity recognition. The present study critically evaluates state-of-the-art (SOTA) techniques in group activity recognition, categorizing existing literature into distinct sub-domains. Furthermore, the paper identifies available datasets and highlights prevailing research challenges within the realm of group activity recognition.